Google's Threat Intelligence Group says a suspected Russian-speaking operator ran what it describes as the first observed instance of an operator using AI to turn security patches into working exploits without human involvement, via an AI framework that found flaws and then built and executed exploits automatically. For UK buyers, this is the clearest signal yet that vulnerability management programmes need a human checkpoint, not more automation for its own sake.
View the data behind this chart
| Cyber Shield | National Exposure Centre | |
|---|---|---|
| Funding | £60m | £70m |
Google finds the first patch-to-exploit chain with no human involved
In its fourth-quarter 2026 AI threat tracker, Google's Threat Intelligence Group (GTIG) logged the incident as a marked leap forward from the agentic activity it had tracked only three months earlier. Investigators located an unsecured staging server running a multi-agent vulnerability-research and weaponisation toolkit — named exploit_agent — built to plug into several frontier and locally hosted AI models, which used Russian-language tooling.
Built for continuous differential analysis of 'popular web application software,' the system compared pre- and post-patch code to pinpoint what a fix had actually addressed, then developed and automatically executed exploits against that flaw, with no human input at any point in the chain.
GTIG's investigation found that the server also contained data stolen from victims, alongside multiple post-compromise scripts capable of internal reconnaissance, brute-force attacks, deploying webshells, and stripping out webshells left by rival intruders. Victim telemetry pointed to a particular focus on French sites tied to media, entertainment and the education space.
Part of a wider shift to agentic offensive frameworks
GTIG frames the case as part of a larger trend: attackers abandoning manual chatbot prompting in favour of self-directed, multi-agent offensive toolkits capable of running intrusions through several stages with minimal supervision.
Beyond the open-source penetration-testing tool ARTEX, GTIG points to a roster of other AI-driven offensive toolkits observed in attacks or threat-actor experimentation — CAIRN, CyberStrikeAI, HexStrike AI and PentAGI among them. Separately, GTIG saw a group combine PentAGI with proxy infrastructure in a campaign against aerospace targets.
"We have observed several threat actors separately experimenting with use of these frameworks and harnesses, including financially motivated, as well as suspected espionage threat actors with a PRC nexus," GTIG's researchers wrote, cautioning that how easily attackers can get started and bolt on fresh exploits means these frameworks are likely to keep appearing in campaigns against organisations worldwide.
The NCSC's warnings now look less theoretical
This lands squarely in territory the National Cyber Security Centre has been flagging. The NCSC assesses that AI is likely to increase the speed, scale and efficiency of existing cyberattacks, while judging that human operators will still be highly likely to remain necessary for directing, validating and adapting AI-generated activity — the gap exploit_agent appears designed to close.
More pointedly, the NCSC has assessed that AI-powered malware is highly likely to feature in a ransomware attack on UK critical national infrastructure by 2028, alongside more effective social engineering and vulnerability exploitation. It has also urged organisations to prepare for what this article terms a 'patch wave' — a rush of updates across the stack following new disclosures — and to move away from periodic patch cycles towards faster, more frequent deployment.
The NCSC says organisations will need greater automation to keep pace, while warning that automation cannot compensate for insecure foundations — a reminder that proactive patch management discipline still matters more than ever.

Why auditable automation, not maximum automation, is the buyer test
The practical requirement emerging for UK infrastructure buyers is not to automate everything, but to make every automated action auditable: organisations need to be able to identify what an AI agent changed, what evidence it relied on, what permissions it held, and which human approved or overrode the action. That follows directly from the NCSC's call for appropriate human oversight and safeguards around AI-enabled defence.
The NCSC's capability model for agentic security tools describes systems that can select tools, sequence actions, inspect results and adapt their approach on the fly — though this describes a capability framework rather than a claim that all deployed commercial products already operate this way. The NCSC says such systems require varying degrees of human oversight, which turns access controls, approval gates and rollback capability into material procurement criteria, not nice-to-haves.
The government has also funded related programmes: Project DROMOS reportedly began as a four-week pilot analysing government code repositories, said to have identified hundreds of potential security issues including critical vulnerabilities, before expanding into operational environments. The follow-on Cyber Shield programme is reported to involve deploying defensive agents across participating organisations, backed by a reported £60 million, alongside a separate reported £70 million for the National Exposure and Real-Time Vulnerability Centre. These programmes are relevant examples of defensive automation investment, though their exact scope and whether they implement auditable human-approval or rollback controls has not been independently confirmed.
View the data behind this chart
| Phase | Starts (week) | Duration (weeks) |
|---|---|---|
| DROMOS pilot | 0 | 4 |
| Cyber Shield rollout | 4 | 152 |
What UK infrastructure buyers should do now
Competitive defence now depends on cutting patch latency without giving up control. Automated prioritisation and testing can accelerate the response to a disclosed flaw, but humans should retain authority over production changes, high-impact containment decisions and exceptions — exactly the model GTIG's findings argue attackers no longer need on the offensive side.
Buyers should benchmark how fast they can move from disclosure to deployed fix using tools like Servnet's CVE exploitation speed index, since GTIG's findings show the exploit_agent framework was configured to compare pre- and post-update code and automate exploit development and execution. It is also worth reviewing exposure to autonomous post-exploitation attacks, which extend the same agentic model into the stages after initial compromise — reconnaissance, lateral movement and data theft.
Segmentation and layered monitoring remain the backstop: organisations assessing zero trust architecture and managed detection & response coverage should treat this disclosure as confirmation that assumptions about attacker speed and sophistication need revisiting now, not at the next budget cycle.
- 01The Register — Google: AI agent turned a security patch into a working exploit with no human help · 16 November 2026
- 02NCSC — AI impact on the cyber threat (Annual Review 2026) · 1 November 2026
- 03The Register — Britain on notice: AI-powered ransomware attacks to arrive by 2028 · 3 November 2026
- 04NCSC Annual Review 2026 (PDF) · 1 November 2026
- 05NCSC — Cyber chief: UK facing AI-powered attacks on critical infrastructure in two years · 1 November 2026
- 06NCSC — Supporting AI adoption for UK cyber defence · 1 November 2026
- 07NCSC — Impact of AI on the cyber threat: now to 2027 · 1 November 2026
- 08NCSC — The changing nature of operational cyber defence · 1 November 2026
